Commerce Protocols for AI Agents: UCP, ACP, AP2, and MCP — How the Stack Fits Together

Key takeaways

• UCP has 11 co-developers including Google, Shopify, Amazon, Meta, and Stripe — it is expanding from shopping into lodging and food, making it the broadest commerce-protocol coalition as of August 2026.
• ACP is governed by OpenAI and Stripe (now joined by Meta), in beta, and live in ChatGPT since September 2025 — it covers agent-driven checkout but not the B2B purchasing workflow.
• AP2 secures agent-led payments across 60+ organizations, with Coinbase x402 as the stablecoin extension — x402 has processed 100 million agentic payments on Coinbase Base as of 2026.
• MCP is the intelligence layer underneath both UCP and ACP — every Shopify UCP store exposes a live MCP endpoint, and ACP's three-layer architecture puts MCP in the middle.
• None of the four protocols covers the B2B semantic layer — customer-tier pricing, bulk RFQ quoting, inventory holds against an ERP, and order attribution across channels all require a custom MCP module on top.

Two open standards launched within months of each other in 2025–2026 to solve the same problem from different directions: how does an AI agent discover a product, negotiate a price, complete a checkout, and pay — without a human wiring up a bespoke integration for every store? The Universal Commerce Protocol (UCP) was co-developed by Google and Shopify with 11 co-developers including Amazon, Meta, Microsoft, Salesforce, and Stripe. The Agentic Commerce Protocol (ACP) is maintained by OpenAI and Stripe and has been live in ChatGPT since September 2025. A third protocol, the Agent Payments Protocol (AP2), secures the payment mandate layer with 60+ participating organizations. All three sit on top of the Model Context Protocol (MCP), which provides the intelligence-layer connection between an agent and a store's data.

This article maps where each protocol fits, how they overlap, and where the gap is for B2B commerce — because the four protocols were designed for consumer discovery and checkout, not for the quoting, tier-pricing, and inventory-hold workflows that a mid-market distributor running NetSuite or BigCommerce actually needs.

The four-layer stack

The commerce protocol landscape is not a single standard. It is a stack of four complementary protocols, each solving a different layer of the agent transaction:

Layer	Protocol	Governed by	What it does	Production status
Discovery & commerce	UCP	Google, Shopify, 11 co-developers	Agent discovers products, builds carts, hands off to merchant for checkout, tracks orders	Live (Shopify Spring '26 Edition); expanding to lodging and food
Checkout interaction	ACP	OpenAI, Stripe, Meta	Agent-driven purchase interaction model; three-layer architecture (Interaction → Intelligence → Commerce)	Beta; live in ChatGPT since Sept 2025
Payment mandate	AP2	Google, 60+ orgs, FIDO Alliance	Verifiable credentials, payment mandates, cryptographic audit trail for agent-led payments	v0.2; FIDO Alliance standardization underway
Intelligence & tools	MCP	Anthropic (originally)	Agent-to-tool connection: reading product data, querying inventory, calling store APIs	Spec final 2026-07-28; ~20K servers registered

MCP is not a commerce protocol — it is the intelligence layer that both UCP and ACP depend on. UCP's specification explicitly states "MCP support built-in." ACP's three-layer architecture places MCP in the middle: the Interaction layer (agent-to-buyer) sits on top, the Intelligence layer (MCP) connects the agent to store data, and the Commerce layer handles fulfillment. Every Shopify store that implements UCP exposes a live MCP endpoint for agents to query product catalogs, manage carts, and fetch order state.

UCP — the discovery-to-checkout standard

UCP is the broadest coalition. The ucp.dev site lists 11 co-developers across three industries: shopping (Google, Shopify, Etsy, Wayfair, Target, Walmart, Amazon, Microsoft, Meta, Salesforce, Stripe), lodging (Amadeus, Booking.com, Expedia, Hilton, Marriott, Trip.com), and food (DoorDash, Square, Toast, Uber Eats). The endorsed partners list includes Visa, Mastercard, Coinbase, PayPal, Adyen, Klarna, and Worldpay.

UCP supports REST and JSON-RPC transports, with AP2, A2A, and MCP support built in. The Shopify developer documentation confirms that Shopify's MCP tools implement UCP at every step of the buyer journey: negotiate and authenticate (profile-based trust tiers), discover products (search across hundreds of millions of listings), carts and checkout (build carts, convert to checkouts, hand off to merchant for payment), and monitor orders (order webhooks and on-demand order state via the get_order MCP tool).

The Universal Cart API lets an AI agent collect items from any merchant — on or off Shopify — into a single unified cart via UCP. Shopify's certification program (Shopify + OpenAI + Google, closed beta, Q3 2026 broader access) creates what amounts to a new discovery layer for agents: clean structured data determines whether an agent can find and transact with a store.

ACP — the checkout interaction model

ACP takes a different architectural approach. Rather than a broad multi-industry coalition, it is governed by OpenAI and Stripe (Stripe's documentation now lists Meta as a co-creator) as Founding Maintainers, with a path toward broader community governance. The spec is in beta under Apache 2.0.

ACP defines composable building blocks for agent-driven commerce. Stripe's Agentic Commerce Suite provides the reference implementation: agents discover products, add them to carts, and complete purchases using Stripe's payment infrastructure. BigCommerce adopted the ACP path via its Stripe partnership rather than building a first-party store-data MCP server — the BigCommerce connector article covers this in depth.

The three-layer ACP architecture — Interaction → Intelligence (MCP) → Commerce — means MCP is the connective tissue. An agent does not bypass MCP to use ACP; it uses MCP to talk to the store and ACP to structure the purchase interaction.

AP2 — the payment mandate layer

AP2 sits underneath both UCP and ACP as the payment-trust protocol. Google announced AP2 in September 2025 with 60+ partner organizations. It is now being standardized within the FIDO Alliance. AP2 extends A2A (agent-to-agent communication) with structured payment mandates — Intent Mandates, Cart Mandates, and Payment Mandates — that provide verifiable, non-repudiable proof that a user authorized an agent to make a specific purchase.

Coinbase x402 is the stablecoin extension to AP2, named after HTTP status code 402 ("Payment Required"). x402 lets agents pay for API calls, services, and micro-transactions directly with stablecoins over HTTP. Chainalysis reported that x402 crossed 100 million agentic payments on Coinbase Base, demonstrating that the stablecoin rail is operating at production scale.

Three competing payment networks have formed around agent commerce: Visa Trusted Agent, Mastercard Agent Pay (30+ industry partners, Agentic Tokens via MDES), and Coinbase x402. Stripe is provisioning agentic network tokens from both Mastercard and Visa, positioning itself as the bridge between traditional card rails and the agent-payment layer.

The four-layer commerce protocol stack, with the B2B semantic-layer gap that requires a custom MCP module:

Where the stack converges — and where it does not

The four protocols are converging on a shared architecture: UCP or ACP for the commerce interaction, AP2 for the payment mandate, MCP for the intelligence layer, and A2A for agent-to-agent delegation. The UCP specification explicitly references AP2, A2A, and MCP as built-in integrations. The AP2 documentation explicitly references A2A and UCP. The stack is designed to interoperate.

But the convergence is consumer-shaped. The protocols solve a specific problem: an agent discovers a product, negotiates a price, completes a checkout, and pays. That is the consumer journey. A mid-market B2B distributor quoting 200 RFQs a week does not have a checkout problem — it has a pricing-tier problem, an inventory-hold problem, and an ERP-write-back problem.

The gap is the B2B semantic layer:

• Customer-tier pricing. UCP and ACP expose a store's published prices. A B2B distributor's prices are per-customer, per-contract, per-volume-tier — stored in NetSuite or BigCommerce customer groups, not in a public catalog. No commerce protocol negotiates tier pricing.
• Bulk RFQ quoting. The consumer protocols handle single-item purchases. B2B procurement runs on RFQs: a buyer sends a request for 500 units with a delivery window, the supplier responds with a quote, the buyer negotiates. The RFQ engine architecture article covers this workflow in depth.
• Inventory holds against an ERP. A consumer checkout reserves inventory at the moment of purchase. A B2B quote needs an availability hold — a time-limited reservation against NetSuite or Brightpearl inventory that expires if the quote is not accepted. No UCP or ACP tool provides this.
• Order attribution across channels. A B2B order placed by an agent needs to be attributed to the right customer account, the right sales rep, and the right contract in the ERP. The consumer protocols assume a single buyer identity.

A custom MCP module fills this gap. The MCP Module Code Standard defines the structural pattern: typed tool definitions, semantic-layer translations between the agent and the ERP, rate-limit governance, and an audit trail. The module sits on top of the commerce protocols — it does not replace UCP or ACP, it extends them into the B2B workflows that the consumer protocols were not designed to cover.

Related reading

• Connecting an AI Agent to BigCommerce with MCP: What the Stripe Partnership Does Not Solve — the per-connector analysis of BigCommerce's ACP path and the B2B semantic-layer gap
• When an AI Agent Sells on Your Behalf: Connecting Shopify to the B2B Stack — how Shopify's UCP-compliant MCP servers cover the consumer half and where a custom module extends into B2B
• MCP + A2A: The Two Protocols Behind Every Production Agentic AI System — the protocol-stack overview that this article extends into the commerce layer

Representative build

A mid-market industrial distributor running BigCommerce for its online catalog and NetSuite for ERP receives 150 RFQs a week by email. Each RFQ requires a customer-tier price lookup, an inventory availability hold, and a quote that references the customer's contract terms. The consumer commerce protocols — UCP for product discovery, ACP for checkout, AP2 for payment — do not touch any of these steps. A custom MCP module bridges the gap: it exposes typed tools for tier-pricing lookup, availability-hold creation, and quote assembly that an agent calls after discovering products through the standard commerce protocol. The agent handles the RFQ end-to-end — discovery, pricing, hold, quote, human approval — while the MCP module governs the ERP write-back and audit trail. First agent live in 5–8 weeks.

Request a scoped build. One-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.